Hybridized artificial neural network classifiers with a novel feature selection procedure based genetic algorithms and information complexity in credit scoring

Abstract

The credit scoring is a statistical analysis performed by financial institutions to represent the creditworthiness of an individual or small and medium‐sized enterprise. A credit score is a numerical quantity that can be qualified by a rating label showing the potential risk. In order to determine which customers are likely to bring in the most revenue at the exact interest rate and credit limits, the lenders take into accounts these credit scores. In this context, the robust statistical models are inevitable to reduce the number of wrong decisions in the credit evaluation process. Although the machine learning approaches provide superior performance to conventional statistical methods, they are mostly criticized due to the selection of model structure, model complexity, tuning parameters, time consumption in the high‐dimensional and excessive nonlinear cases. For this reason, this study introduces an efficient model estimation and feature selection procedure for artificial neural network (ANN) classifiers in the context of credit scoring. Essentially, this procedure hybridizes training of ANNs with a novel feature selection approach based on genetic algorithms and information complexity criterion. In the application, the proposed procedure was performed on a couple of benchmark credit scoring datasets. According to analysis results, the proposed approach not only estimates robust models from ANNs in terms of model complexity, feature selection, and time consumption, but also outperforms the traditional training procedure for the classification accuracies, false positive, and false negative errors overtraining and test datasets.

Clinical prognosis evaluation of COVID-19 patients: An interpretable hybrid machine learning approach

“…In literature, there exist many approaches in order to reduce the dimension of feature matrix such as all possible subset, forward selection, backward elimination, stepwise selections or some transformation techniques as PCA, ICA, factor analysis, fuzzy relations and their hybrid versions. However, these procedures have some shortcomings, and require some detailed statistical evaluation [ 16 , 20 , 21 , 23 ]. For instance, all possible subset procedure needs 2 p iterations to determine the best feature subset where p is the number of features in the system.…”

“…To overcome this computational burden, the hybrid feature selection procedures based evolutionary algorithms (such as GAs, PSO, Bee and Ant Colonies) and selection criteria are mostly preferred. In the literature, there exist remarkable studies in which GAs are used for feature selection procedure [ 21 , [23] , [24] , [25] , [26] ]. For this reason, in this study, the feature selection procedure is based on GAs.…”

“…However, let's look at the previous studies in literature, then it can be seen that the model complexity is often overlooked by the analysts. As a result of this, the model complexity causes two type errors: the approximation and estimation known as variance and bias trade-off as well [19] , [20] , [21] . In the training procedure, the model complexity is controlled by means of various approaches such early stopping procedure with cross-validation, regularization, feature selection, information criteria, etc.…”

“…In the training procedure, the model complexity is controlled by means of various approaches such early stopping procedure with cross-validation, regularization, feature selection, information criteria, etc. Essentially, using information criteria in modeling by ML methods provides a crucial role to control the model complexity in terms of constructing an appropriate approximated model, reducing the feature size and choosing the best model in the alternatives [ 21 , 22 ].…”

“…According to the statistical learning jargon, Occam's razor (also called as the law of parsimony) provides an inevitable framework to control the model complexity, so it is widely considered in regression, classification, clustering and time series problems. However, this rule doesn't work properly to evaluate the best fitting models in the classification tasks, because Akaike-type criteria (AIC, BIC, CAIC, DIC) do not deal with the discriminative information regarding to competing classes in the classification problems [ 16 , 21 ]. For this reason, discriminative based information criteria such as ICOMP are more convenient for the classification task.…”